Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
  • Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
  • Tesla's Supercharger Team was recently laid off. We discuss what this means for the company on today's TMC Podcast streaming live at 1PM PDT. You can watch on X or on YouTube where you can participate in the live chat.
This site may earn commission on affiliate links.
The next big milestone for FSD is 11. It is a significant upgrade and fundamental changes to several parts of the FSD stack including totally new way to train the perception NN.

From AI day and Lex Fridman interview we have a good sense of what might be included.

- Object permanence both temporal and spatial
- Moving from “bag of points” to objects in NN
- Creating a 3D vector representation of the environment all in NN
- Planner optimization using NN / Monte Carlo Tree Search (MCTS)
- Change from processed images to “photon count” / raw image
- Change from single image perception to surround video
- Merging of city, highway and parking lot stacks a.k.a. Single Stack

Lex Fridman Interview of Elon. Starting with FSD related topics.


Here is a detailed explanation of Beta 11 in "layman's language" by James Douma, interview done after Lex Podcast.


Here is the AI Day explanation by in 4 parts.


screenshot-teslamotorsclub.com-2022.01.26-21_30_17.png


Here is a useful blog post asking a few questions to Tesla about AI day. The useful part comes in comparison of Tesla's methods with Waymo and others (detailed papers linked).

 
Last edited:
A lot of stuff we don't know but the one thing we do is that the car could self-park when it was shipping with USS and cannot do it now (I've posted a couple of times on this, won't do it again now). I get it that self-driving has never been done before, and it will be super impressive if Tesla gets there with just the cameras, but parking sensors, rain sensors, etc - all this has been done before. Why not keep this tried and tested - and cheap - tech until you have a vision-only solution?
Because Tesla doesn't care. They can keep kicking the can down the road, hoping we all sell our cars before any goodwill toward them runs out. I can guarantee someone (many) at Tesla knows FSD is never coming in any truly meaningful way (i.e. beyond very, very basic Level 2) to existing cars on the road. Robotaxis, for example, with current models = never happening.
 
All right. I've been half-following this bit about the "B pillar blocks the view" on and off for a bit. And something's been bothering me about it: I thought that there were overlapping camera views in pretty much every direction.

So, just did a short search and, NotATeslaApp had this:
1701099656785.png

So, yeah, they've got ranges plonked down, presumably based upon the pixel resolution of all the cameras. But, you'll notice, there's a 360-degree view in all blame directions. And where one camera stops looking, there's at least one other camera with overlap.

Yeah, everybody knows that those, "repeater" cameras only look towards the side/rear: But there are cameras that look straight out sideways (light green) which, given the above picture posted above by @aronth5, kind of covers, fully, that hydrant and the auto coming up the road.

What am I missing, here? Is somebody claiming that some cameras aren't being used for FSD?
 
....But there are cameras that look straight out sideways (light green) which, given the above picture posted above by @aronth5, kind of covers, fully, that hydrant and the auto coming up the road.

What am I missing, here? Is somebody claiming that some cameras aren't being used for FSD?
First that is just older info off of the Tesla's site. Also the Light Green is from the B-Pillar camera that is blocked from "seeing" the hydrant in the pic.
 
Because Tesla doesn't care. They can keep kicking the can down the road, hoping we all sell our cars before any goodwill toward them runs out. I can guarantee someone (many) at Tesla knows FSD is never coming in any truly meaningful way (i.e. beyond very, very basic Level 2) to existing cars on the road. Robotaxis, for example, with current models = never happening.
Even the L2 they offer is more unreliable/predictable than 10 year old L2 offered by others - Apples to Apples.
 
By the fence. (The cameras don't have x-ray vision.)
Ah. So this is a case of "different angles of view".

Back in the day, the SO and I owned a 1986 Toyota Camry. They were very proud of the fact that the A-pillars (Hope I got this right, it's the ones on the front corners of the windshield) were narrower than the distance between a human's pair of eyeballs. So that when one looked towards the pillar, on one side or the other, one would see what was on the other side of the pillar, since one eye or the other would have a view.

Cute, but every car the SO or I have owned since then have had thicker A pillars. Not knowing for sure, but I'd guess that those suckers have all been thicker so the car's passenger compartment could survive rollover tests.

Having said that, there's a decent argument that, with a human driver, there's a naturally occurring blind spot to the left and right where, without bobbing the human's head back and forth, Nothing Can Be Seen.

Along those lines, though, the distance to said blocked object, the size of said blocked object, the movement of said blocked object, and the movement of one's own car (including linear as well as rotational) will all come into play. As a worst case example, there's some speeder flying down towards one that isn't immediately viewable; then, as one turns to the left, the left A-pillar tracks right along with the speeder, making for a Surprise! when it gets close enough to be wider than the pillar. That's pretty obviously a corner case: Humans are fairly good at detecting moving objects coming in and out of view and all that.

Even with all that.. There's an intersection near here where a building is built up to the road coming up from the left; there's no room for a sidewalk and about six inches from the curb to the building front. The traffic control is a stop sign. Never mind the white line stuff: One has to stick the car's nose out into the road in order to see approaching traffic, which is dangerous, because that traffic is moving.

So, even for humans, some blockage of vision is a Problem.

At least with the Tesla cameras, whether or not said cameras are a bit farther back than a human's head, there's no issues with the A pillar getting in the way. So that's an improvement. If I'm following this argument correctly, what's being requested is a left/right-view cameras planted on the A pillars, farther up than the driver, and so having bit better views.
 
  • Informative
Reactions: FSDtester#1
All right. I've been half-following this bit about the "B pillar blocks the view" on and off for a bit. And something's been bothering me about it: I thought that there were overlapping camera views in pretty much every direction.
[... with the image of the overlapping views ...]
What am I missing, here? Is somebody claiming that some cameras aren't being used for FSD?

The difference is that each of the cameras comes from a fixed point of view on the car, while a human's head can move. When you think about obstructions (like the fence example with pics just earlier in this thread). There will always be obstructions that block your (or the car's) view in ways that matter for safety, even if you can see in 360 degrees from a single point. Humans solve this, in many cases, by craning their head around to get a slightly different point of view (e.g. leaning out over the steering wheel).

The way you fix this for ADAS cameras is that you need critical areas covered by more than one camera from more than one point of view. The B-pillar is the most important example. In your diagram, the yellow areas on the front sides are only covered by the B-pillar camera, and they cover a pretty critical area for safety checks when turning. If there was another camera covering the yellow areas from a different POV, we'd be fine. Imagine if the rear-angle-facing repeaters on the front fenders were wider-angled and could see straight sideways as well. Now you'd have two independent points-of-view into those critical yellow areas, and any reasonable obstruction blocking one camera from seeing a car there probably wouldn't block the other. Having just one or the other (e.g. just the sideways-seeing repeaters, but delete the b-pillars) will always be subject to some scenario where safety is compromised by a trivial obstruction.

On the other other hand: IMHO in many cases, creep behavior into shoulder/crosswalk areas and/or angling (e.g. turning to the right while creeping, to angle the repeater out to the left), should be able to solve this problem as well. Those cases which can't be fixed in that way are honestly terribly unsafe intersection designs to begin with, even for humans. Still, good luck getting every city to fix every such problem for you. The real world is a bitch.
 
Those cases which can't be fixed in that way are honestly terribly unsafe intersection designs to begin with, even for humans. Still, good luck getting every city to fix every such problem for you.
What's next? Blaming stupid drivers for not understanding roundabouts? Unfortunately, the RW road network is laced with poorly designed intersections. Tesla needs to compensate by adding either additional cameras or sensors. Else, the only option is to have the navigation algorithm avoid dangerous intersections.
 
Here's another tough intersection where FSD rolls the dice on an UPL. The UI interface says 'creeping forward for visibility' but FSD appears committed to making the turn based on poor visibility from moments earlier.

Funny. I've had issues like this before.

Very near to my place there's a bog-standard intersection where one road T's into a bit more major road. Only choices are left and right and there's a stop sign, complete with white line. 90 degrees, clear views to the right and left.

Roughly 85% of the time the car will halt at the white line, creep up a bit for the UPL, wait for cross-traffic to go away, then go for it.

Roughly 15% of the time it'll try the business with the video: Not see approaching cars for some reason and go for it anyway, from the white line without creeping up first.

Conclusion: Not consistent. Dunno if this is camera issues or Something Else.
 
Funny. I've had issues like this before.

Very near to my place there's a bog-standard intersection where one road T's into a bit more major road. Only choices are left and right and there's a stop sign, complete with white line. 90 degrees, clear views to the right and left.

Roughly 85% of the time the car will halt at the white line, creep up a bit for the UPL, wait for cross-traffic to go away, then go for it.

Roughly 15% of the time it'll try the business with the video: Not see approaching cars for some reason and go for it anyway, from the white line without creeping up first.

Conclusion: Not consistent. Dunno if this is camera issues or Something Else.
Perhaps FSDb does not distinguish between not seeing another car and seeing that there is not another car.
 
At least with the Tesla cameras, whether or not said cameras are a bit farther back than a human's head, there's no issues with the A pillar getting in the way. So that's an improvement. If I'm following this argument correctly, what's being requested is a left/right-view cameras planted on the A pillars, farther up than the driver, and so having bit better views.
Just left/right cameras that are closer to the front, not 9 feet behind the front of the car. It's all about improving FSD visibility when there are obstructions limited how well FSD can see cross traffic. Whenever a human has to lean forward visibility relying on the B-pillar camera is compromised. Of course if FSD can creep far enough then your ok but sometimes that is not an option. So these are edge cases for sure but FSD is supposed to be much safer than a human driver.
Lots of choices,
  • A-pillar,
  • Add to existing fender cameras
  • Add to front windshield location which has the advantage of being protected from weather plus higher is better
  • Top roof above windshield location
  • Headlight area
 
  • Like
Reactions: FSDtester#1
First road trip (1000 miles round trip) with FSD and overall I was really happy with it. I know it's basically autopilot on the highway -- but I have found FSD to be better than AP on highways with the exception of crazy lane changes. I was pretty worried about phantom braking and crazy lane changes but lucked out with zero phantom braking and found that the minimal lane change setting worked great. The car drove super smoothly and I loved having the car make the lane changes for me when I signaled. It's amazing how little things like that can make a difference in fatigue driving.

I went into the trip with enough experience with FSD to be be wary but ended the trip really loving it. At first I was annoyed by the constant nagging but I grew to appreciate the fact that it did indeed keep me focused on the road. I kept telling myself that I am driving the car and that FSD is just a really good drive assistance program. Still, it's easy to get complacent so the "annoying" nagging was probably okay in the end.

Here are a few of the issues I did have:
1. It doesn't handle cross traffic on highways well. A car can cross the highway in plenty of time but my car would slow down a lot even though there was no longer a car in front of me. I don't count this as phantom braking as I can see what's causing it and it's really easy to account for (I was on alert anyway as I figured this could be a problem). Not a big deal as this doesn't happen a lot and again it's easy to just keep the foot on the accelerator when this happens.
2. Frequent warnings of camera degradation. Not sure what causes this but during certain stretches this warning would come and go. I kept the system on but was just more ready to take over. The system worked fine but the beeping every minute was a bit annoying (though better than not alerting for sure). The problem would clear itself up in five minutes or so, so I'm not sure what was really going on.
3. It did not take exits well at all. It would exit a little late and not signal until after making the move. It also would take them at a worryingly high speed. I really don't need the car to do this on it's own so not a a big deal, but I was kind of suprised.
4. One time the car swerved into a turn lane on the right instead of going straight -- I assume, because of the weird markings of the lane. Not a big deal as I was able to correct. Frankly, I'm surprised it only had one lane keeping issue in over 1000 miles.


Overall, it was pretty amazing and made the trip a lot easier. It's always good when you appreciate your car more after a road trip.
 
Here's another tough intersection where FSD rolls the dice on an UPL. The UI interface says 'creeping forward for visibility' but FSD appears committed to making the turn based on poor visibility from moments earlier.

Starting at 2:35 the B-pillar's view is restricted by the bushes/tree. Not a lot but just enough that the driver felt very unsafe when the car pulled out. Classic problem with FSD.
 
2. Frequent warnings of camera degradation. Not sure what causes this but during certain stretches this warning would come and go. I kept the system on but was just more ready to take over. The system worked fine but the beeping every minute was a bit annoying (though better than not alerting for sure). The problem would clear itself up in five minutes or so, so I'm not sure what was really going on.
A couple cases where I get this message are:
  • Cold weather soon after start driving. Condensation can build up inside one, or more, camera housings that fogs up the image. After a while, the area warms up and the condensation evaporates.
  • Night driving in rural areas where there are stretches with no lights to one, or both, sides of the car. It's simply so dark that the side cameras don't see anything and FSD assumes that the camera is obscured.
 
A couple cases where I get this message are:
  • Cold weather soon after start driving. Condensation can build up inside one, or more, camera housings that fogs up the image. After a while, the area warms up and the condensation evaporates.
  • Night driving in rural areas where there are stretches with no lights to one, or both, sides of the car. It's simply so dark that the side cameras don't see anything and FSD assumes that the camera is obscured.
I get the B-Pillar cameras fogging up in the winter when sunlight hits em. My understanding is that HW4 solves this with heaters in the B-Pillar housing. If that's true, it'll help quite a bit for the new cars.
 
Lots of choices,
  • A-pillar,
  • Add to existing fender cameras
  • Add to front windshield location which has the advantage of being protected from weather plus higher is better
  • Top roof above windshield location
  • Headlight area
I agree with most of those, I still think the headlights, with the appropriate optical glare isolation design, would make a lot of sense.

One location that you didn't put in your list is at or near the mirrors. I believe MobilEye is using that location ( but I still like the headlights.

This could have been instead of, not in addition to, one or two of the existing front cameras. But I think it's water under the bridge now.
 
I agree with most of those, I still think the headlights, with the appropriate optical glare isolation design, would make a lot of sense.

One location that you didn't put in your list is at or near the mirrors. I believe MobilEye is using that location ( but I still like the headlights.

This could have been instead of, not in addition to, one or two of the existing front cameras. But I think it's water under the bridge now.
My preference for HW5 is adding two cameras with 90 degree wide angle view to the existing windshield area for the following reasons: Not sure though if they could be added in such a way to avoid sun glare. The key is that every time a human leans forward to monitor FSD decisions at intersections due to any type of obstruction means FSD does not see as well as the human.
  1. Protected by rain/snow since the wipers/defrost clear the windshield
  2. Cameras are high so they will see over vehicle hoods and snow banks
  3. They are far enough forward to be a real improvement over the human driver view, mirrors and A-pillar.
 
  • Like
Reactions: FSDtester#1